PAT 04
IMPLEMENTATION PATTERN
Versioned evaluation fixture
Prompt and model changes are judged from memory or a single favorable example.
Any behavior change can affect quality, safety, review effort, latency, or cost.
REQUIRED CONTRACT
Keep these visible.
- 01Representative cases
- 02Expected behavior
- 03Observable rubric
- 04Frozen configuration
AVOID
Do not hide the failure.
- One average score
- Unlabeled synthetic data
- Moving expectations
- No failure cases
Prove the boundary.
Change one material configuration and compare the same cases, reviewer minutes, failures, latency, and cost units.